Hi all, I created a repository to share the way how I trained the model I used in this notebook
\n\nI think the weight file I shared here can be reproduced using the script I share here.
\n\nSorry for the delay, I at first tried on kaggle notebook but couldn't make a training notebook because of the limitation of kaggle dataset.
\n\nI hope this will give you a good insight and eventually lead you to a progress.\nHave happy kaggling days!
", "messages": [ { "id": "905227", "postDate": "06/28/2020 11:19:30", "content": "Hi all, I created a repository to share the way how I trained the model I used in this notebook
\n\nI think the weight file I shared here can be reproduced using the script I share here.
\n\nSorry for the delay, I at first tried on kaggle notebook but couldn't make a training notebook because of the limitation of kaggle dataset.
\n\nI hope this will give you a good insight and eventually lead you to a progress.\nHave happy kaggling days!
", "rawMarkdown": "Hi all, I created [a repository](https://github.com/koukyo1994/kaggle-birdcall-resnet-baseline-training) to share the way how I trained the model I used in [this notebook](https://www.kaggle.com/hidehisaarai1213/inference-pytorch-birdcall-resnet-baseline)\n\nI think the weight file I shared [here](https://www.kaggle.com/hidehisaarai1213/birdcall-resnet50-init-weights) can be reproduced using the script I share here.\n\nSorry for the delay, I at first tried on kaggle notebook but couldn't make a training notebook because of the limitation of kaggle dataset.\n\nI hope this will give you a good insight and eventually lead you to a progress.\nHave happy kaggling days!", "votes": null }, { "id": "916364", "postDate": "07/05/2020 15:25:15", "content": "Great work! Upvoted all your notebooks and dataset
", "rawMarkdown": "Great work! Upvoted all your notebooks and dataset", "votes": null }, { "id": "917410", "postDate": "07/06/2020 13:31:36", "content": "That's amazing, it makes it really hard to get started!
\n\nIn terms of workflow, would you then usually download the inputs and work locally rather than within a remote kaggle notebook?
\n\nI'd imagine it depends on your local set up and network throughput? pr🙏
", "rawMarkdown": "That's amazing, it makes it really hard to get started!\n\nIn terms of workflow, would you then usually download the inputs and work locally rather than within a remote kaggle notebook?\n\nI'd imagine it depends on your local set up and network throughput? pr🙏", "votes": null }, { "id": "920393", "postDate": "07/08/2020 15:02:53", "content": "Yes, I downloaded the data to local environment and did all the experiment on it.
\n\nIn fact, I have access to some of V100s in DGX thus I did experiment on them, network throughput is around 20~50MB/s for download and 10~20MB/s for upload I think although I have not checked so much.
", "rawMarkdown": "Yes, I downloaded the data to local environment and did all the experiment on it.\n\nIn fact, I have access to some of V100s in DGX thus I did experiment on them, network throughput is around 20~50MB/s for download and 10~20MB/s for upload I think although I have not checked so much.", "votes": null }, { "id": "953318", "postDate": "07/31/2020 17:35:27", "content": "Is it possible to train a model with in kaggle environment on the given dataset by passing limited batch to the training loop ?
", "rawMarkdown": "Is it possible to train a model with in kaggle environment on the given dataset by passing limited batch to the training loop ?", "votes": null }, { "id": "953637", "postDate": "08/01/2020 00:09:38", "content": "Yes, that's possible. See @ttahara 's notebook. It's based on my repository and changed some like the backbone model or the training framework (Catalyst -> pytorch-pfn-extras)
", "rawMarkdown": "Yes, that's possible. See @ttahara 's [notebook](https://www.kaggle.com/ttahara/training-birdsong-baseline-resnest50-fast). It's based on my repository and changed some like the backbone model or the training framework (Catalyst -> pytorch-pfn-extras)", "votes": null }, { "id": "953813", "postDate": "08/01/2020 05:57:14", "content": "Yeah thanks for the response @hidehisaarai1213 why everone submitting the code only based on RESNET Model, other model does not working for this dataset?
", "rawMarkdown": "Yeah thanks for the response @hidehisaarai1213 why everone submitting the code only based on RESNET Model, other model does not working for this dataset?", "votes": null }, { "id": "953822", "postDate": "08/01/2020 06:23:19", "content": "I don't think so, it's just a baseline. The point is, you can use larger model, but only enlarging the model will not solve the true challenge in this competition.
", "rawMarkdown": "I don't think so, it's just a baseline. The point is, you can use larger model, but only enlarging the model will not solve the true challenge in this competition.", "votes": null }, { "id": "954088", "postDate": "08/01/2020 11:19:25", "content": "Thanks man! Is there any reason that you have chosen a sampling rate of 32khz? Are bird calls normally below 16khz?
", "rawMarkdown": "Thanks man! Is there any reason that you have chosen a sampling rate of 32khz? Are bird calls normally below 16khz?", "votes": null }, { "id": "954154", "postDate": "08/01/2020 12:52:16", "content": "The host suggested so.
", "rawMarkdown": "The host suggested so.", "votes": null }, { "id": "954223", "postDate": "08/01/2020 14:11:15", "content": "Thank for the information , As there are many bird voices in single file we need to git rid of that during the training process and then train a model,so that we can get good accuracy
", "rawMarkdown": "Thank for the information , As there are many bird voices in single file we need to git rid of that during the training process and then train a model,so that we can get good accuracy", "votes": null } ], "comments": [ { "id": 916364, "author_name": "salmaneunus", "author_url": "", "post_date": "07/05/2020 15:25:15", "content": "Great work! Upvoted all your notebooks and dataset
", "votes": null, "replies": [] }, { "id": 917410, "author_name": "ratsimihah", "author_url": "", "post_date": "07/06/2020 13:31:36", "content": "That's amazing, it makes it really hard to get started!
\n\nIn terms of workflow, would you then usually download the inputs and work locally rather than within a remote kaggle notebook?
\n\nI'd imagine it depends on your local set up and network throughput? pr🙏
", "votes": null, "replies": [ { "id": 920393, "author_name": "hidehisaarai1213", "author_url": "", "post_date": "07/08/2020 15:02:53", "content": "Yes, I downloaded the data to local environment and did all the experiment on it.
\n\nIn fact, I have access to some of V100s in DGX thus I did experiment on them, network throughput is around 20~50MB/s for download and 10~20MB/s for upload I think although I have not checked so much.
", "votes": null, "replies": [] } ] }, { "id": 953318, "author_name": "vpkprasanna", "author_url": "", "post_date": "07/31/2020 17:35:27", "content": "Is it possible to train a model with in kaggle environment on the given dataset by passing limited batch to the training loop ?
", "votes": null, "replies": [ { "id": 953637, "author_name": "hidehisaarai1213", "author_url": "", "post_date": "08/01/2020 00:09:38", "content": "Yes, that's possible. See @ttahara 's notebook. It's based on my repository and changed some like the backbone model or the training framework (Catalyst -> pytorch-pfn-extras)
", "votes": null, "replies": [] }, { "id": 953813, "author_name": "vpkprasanna", "author_url": "", "post_date": "08/01/2020 05:57:14", "content": "Yeah thanks for the response @hidehisaarai1213 why everone submitting the code only based on RESNET Model, other model does not working for this dataset?
", "votes": null, "replies": [] }, { "id": 953822, "author_name": "hidehisaarai1213", "author_url": "", "post_date": "08/01/2020 06:23:19", "content": "I don't think so, it's just a baseline. The point is, you can use larger model, but only enlarging the model will not solve the true challenge in this competition.
", "votes": null, "replies": [] }, { "id": 954223, "author_name": "vpkprasanna", "author_url": "", "post_date": "08/01/2020 14:11:15", "content": "Thank for the information , As there are many bird voices in single file we need to git rid of that during the training process and then train a model,so that we can get good accuracy
", "votes": null, "replies": [] } ] }, { "id": 954088, "author_name": "dunky11", "author_url": "", "post_date": "08/01/2020 11:19:25", "content": "Thanks man! Is there any reason that you have chosen a sampling rate of 32khz? Are bird calls normally below 16khz?
", "votes": null, "replies": [ { "id": 954154, "author_name": "hidehisaarai1213", "author_url": "", "post_date": "08/01/2020 12:52:16", "content": "The host suggested so.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "905227": "Hi all, I created [a repository](https://github.com/koukyo1994/kaggle-birdcall-resnet-baseline-training) to share the way how I trained the model I used in [this notebook](https://www.kaggle.com/hidehisaarai1213/inference-pytorch-birdcall-resnet-baseline)\n\nI think the weight file I shared [here](https://www.kaggle.com/hidehisaarai1213/birdcall-resnet50-init-weights) can be reproduced using the script I share here.\n\nSorry for the delay, I at first tried on kaggle notebook but couldn't make a training notebook because of the limitation of kaggle dataset.\n\nI hope this will give you a good insight and eventually lead you to a progress.\nHave happy kaggling days!", "916364": "Great work! Upvoted all your notebooks and dataset", "917410": "That's amazing, it makes it really hard to get started!\n\nIn terms of workflow, would you then usually download the inputs and work locally rather than within a remote kaggle notebook?\n\nI'd imagine it depends on your local set up and network throughput? pr🙏", "920393": "Yes, I downloaded the data to local environment and did all the experiment on it.\n\nIn fact, I have access to some of V100s in DGX thus I did experiment on them, network throughput is around 20~50MB/s for download and 10~20MB/s for upload I think although I have not checked so much.", "953318": "Is it possible to train a model with in kaggle environment on the given dataset by passing limited batch to the training loop ?", "953637": "Yes, that's possible. See @ttahara 's [notebook](https://www.kaggle.com/ttahara/training-birdsong-baseline-resnest50-fast). It's based on my repository and changed some like the backbone model or the training framework (Catalyst -> pytorch-pfn-extras)", "953813": "Yeah thanks for the response @hidehisaarai1213 why everone submitting the code only based on RESNET Model, other model does not working for this dataset?", "953822": "I don't think so, it's just a baseline. The point is, you can use larger model, but only enlarging the model will not solve the true challenge in this competition.", "954088": "Thanks man! Is there any reason that you have chosen a sampling rate of 32khz? Are bird calls normally below 16khz?", "954154": "The host suggested so.", "954223": "Thank for the information , As there are many bird voices in single file we need to git rid of that during the training process and then train a model,so that we can get good accuracy" }, "source": "meta" }